(B-220) Inverse probability weighting to adjust for informative censoring when estimating Desirability of Outcome Ranking (DOOR) probabilities

Background: The Desirability of Outcome Ranking (DOOR) is a novel composite ranked outcome method created to improve measurement of patient-centered outcomes. It assesses potential combinations or series of outcomes a patient may experience during follow-up, offering greater construct validity than discrete endpoints taken alone. However, methods to account for bias from informative censoring have not been evaluated for use in DOOR.

Objectives: To evaluate the performance of inverse probability of censoring weighted (IPCW) DOOR probabilities to address bias due to informative censoring.

Methods: We simulated 2,000 cohorts each with n=10,000 randomly assigned with p(0.5) to standard (E=0) or new (E=1) treatment. The DOOR had four possible ranks: 1) Alive with no hospitalization; 2) Alive with hospitalization; 3) Dead with no hospitalization; 4) Dead with hospitalization. For simplicity, we assumed hospitalization and death at any time during the uniform period of follow-up were measured once. The effect of treatment was simulated under two scenarios: null (50% DOOR probability) and non-null effect (55% DOOR probability favoring the new treatment). Covariates included continuous age and an age-dependent dichotomous comorbidity risk factor for hospitalization and death. Hospitalization risk increased with age and comorbidity; mortality increased with age, comorbidity, and hospitalization. We simulated censoring under three outcome missingness patterns: missing completely at random (MCAR); missing at random (MAR) conditional on treatment and comorbidity; and missing not at random (MNAR). For each scenario, we estimated crude and IPCW DOOR probabilities. We report the median across 2000 iterations and empirical 95% confidence limits (2.5th to 97.5th percentiles).

Results: Crude DOOR probabilities under MCAR were unbiased for null (49.8% [48.7-50.9]) and non-null scenarios (55.0% [54.0-56.1]) but biased upward under MAR (null: +2.3%; non-null: +2.0%) and MNAR (null: +8.0%; non-null: +6.7%). After IPCW, DOOR probabilities were unbiased under MCAR (null: 49.8% [48.8-50.9]; non-null: 55.0% [54.1-56.1]) and MAR (null: 49.8% [48.7-50.9]; non-null: 55.0% [54.0-56.1]). However, under MNAR, substantial bias remained in both null and non-null scenarios (null: 54.8% [53.7-55.9]; non-null: 59.3% [58.1-60.3]). Precision was not greatly impacted for any missing mechanism (all IPCW 95% CIs had width of 2.0 to 2.2%).

Conclusions: IPC weighting provides a robust method for estimating unbiased DOOR probabilities when covariates that influence informative censoring are measured. Further research is needed to explore performance of this novel outcome measure in more complex settings, such as with additional ranks, tiebreakers, and partially missing outcome data.